Learning agents for iterative voting

Stéphane Airiau (1, 2) , Umberto Grandi (3, 4) , Filipo Studzinski Perotto (5, 6)

1 Université Paris Dauphine-PSL
2 LAMSADE - Laboratoire d'analyse et modélisation de systèmes pour l'aide à la décision
3 IRIT-LILaC - Logique, Interaction, Langue et Calcul
4 UT Capitole - Université Toulouse Capitole
5 IRIT-SMAC - Systèmes Multi-Agents Coopératifs
6 UT3 - Université Toulouse III - Paul Sabatier

This paper assesses the learning capabilities of agents in a situation of collective choice. Each agent is endowed with a private preference concerning a number of alternative candidates, and participates in an iterated plurality election. Agents get rewards depending on the winner of each election, and adjust their voting strategy using reinforcement learning. By conducting extensive simulations, we show that our agents are capable of learning how to take decisions at the level of well-known voting procedures, and that these decisions maintain good choice-theoretic properties when increasing the number of agents or candidates.

Mots clés: Iterative voting; Bandit algorithms; Computational social choice

Domaines: Système multi-agents [cs.MA]

Stéphane Airiau, Umberto Grandi, Filipo Studzinski Perotto. Learning agents for iterative voting. International Conference on Algorithmic Decision Theory (ADT 2017), Oct 2017, Luxembourg, Luxembourg. pp.139-152, ⟨10.1007/978-3-319-67504-6_10⟩. ⟨hal-02641165⟩